Model Card for mmatinm/mpersian_xlm_roberta_large

This model is a fine-tuned XLM-RoBERTa Large for Persian question answering, trained on the PersianQA dataset.
It was originally based on pedramyazdipoor/persian_xlm_roberta_large, which was fine-tuned on the PQuAD dataset.

---

Model Details

Model Description

• Developed by: mmatinm
• Base model: XLM-RoBERTa Large (from Hugging Face Transformers)
• Language(s): Persian (fa)
• Task: Extractive Question Answering (SQuAD v2 style)
• Finetuned from: pedramyazdipoor/persian_xlm_roberta_large

Model Sources

• Repository: GitHub Repo
• Demo: Colab Demo

---

Uses

Direct Use

• Answering questions given a Persian context paragraph.
• Can be used as a QA backend in chatbots or search engines for Persian content.

Downstream Use

• Further fine-tuning for domain-specific QA in Persian.
• Integration into multi-lingual QA systems.

Out-of-Scope Use

• Generative QA (this is extractive only).
• Languages other than Persian.

---

Bias, Risks, and Limitations

• Model performance is dependent on the quality and coverage of PersianQA.
• May fail on highly domain-specific or slang-heavy texts.
• May return incorrect spans for ambiguous questions.

---

How to Get Started with the Model

from transformers import AutoTokenizer, AutoModelForQuestionAnswering

repo_id = "mmatinm/mpersian_xlm_roberta_large"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForQuestionAnswering.from_pretrained(repo_id)

Results

Model & Method	F1 Score	EM Score	No-Answer F1
XLM-R ( LoRA + QA Head)	85.3	71.6	90.7

Citation

@misc{mmatinm2025xlmr,
  title={Persian XLM-RoBERTa Fine-Tuned on PersianQA},
  author={Matin M.},
  year={2025},
  publisher={Hugging Face}
}